An Agent Harness for Enterprises Tired of Rebuilding Orchestration

What it does

CUGA is a Python agent harness aimed at teams building more than a chatbot. It executes tasks across web browsers and APIs, wires in tools via OpenAPI specs or MCP servers, and wraps the whole stack in a policy layer that includes human-in-the-loop approval gates. A built-in management UI lets you draft and publish versioned agent configurations, and the repo includes Helm charts for Kubernetes self-hosting.

The interesting bit

Most open-source agents ship as black-box demos; CUGA ships as a configurable chassis. It separates domain concerns—tools, skills, knowledge bases, and safety policies—from core orchestration, and it backs that modularity with public benchmark claims: #1 on AppWorld and a top-tier WebArena ranking. There is even an experimental “save and reuse” mode that treats prior execution trajectories as cached artifacts rather than disposable logs.

Key highlights

• Hybrid web-and-API execution via Playwright and a browser extension, or pure API mode.
• Five policy types (Intent Guard, Playbook, Tool Approval, Tool Guide, Output Formatter) with human-in-the-loop gates.
• Knowledge engine with RAG over PDFs, Office docs, and Markdown via Docling, scoped to either agent-level or per-session contexts.
• Multi-agent supervision and A2A support, including the ability to expose CUGA itself as a tool to other agents.
• Draft-and-publish workflow for agent configs, with versioned releases and a web-based management UI.

Caveats

• The save-and-reuse feature is experimental, and the README does not detail how it handles stateful or side-effect-laden tasks.
• Advanced capabilities like memory, reflection, and hybrid browser mode require flipping TOML flags and installing optional extras, so the default agent is more stripped-down than the feature list suggests.
• The authors note the project is still “evolving toward enterprise-grade reliability,” which suggests production hardening is a work in progress.

Verdict

Worth a look if you are building internal agents that must hit real APIs, respect guardrails, and run on your own cluster. Pure tinkerers looking for a single-script agent may find the configuration surface area overwhelming.